A Scalable Event Dispatching Library for Linux Network Servers

Transcription

1 A Scalable Event Dispatching Library for Linux Network Servers Hao-Ran Liu and Tien-Fu Chen Dept. of CSIE National Chung Cheng University

2 Traditional server: Multiple Process (MP) server A dedicated process to every connection. Concurrency is provided by OS. Disadvantage Context-switching overhead Synchronization overhead TLB miss rate Example Apache Process 1 Process 2 Kernel Connection 1 Connection 2

3 Modern server: Single Process Event Driven (SPED) server User level concurrency within a single process. Nonblocking I/O A mechanism to detect I/O events Disadvantage Influenced by inefficient design of event dispatching mechanism in OS kernel A single page fault or disk read suspends whole server. Example Squid web cache

4 Event-driven server architecture User thread context or Connection processing context Single process Kernel Connections

5 Problem statement and our goal The problem Inefficiency design of event dispatching mechanism Our goal Improve the performance of SPED servers in user-mode.

6 A list of event dispatching mechanisms Scalable Devpoll (Solaris 8) RT signals (Linux) I/O completion port (Windows 2000) Non-scalable select() (POSIX) poll() (POSIX)

7 poll() and select() Event dispatching mechanisms in Linux select() or poll() scales poorly with large set of file descriptors. 30% of CPU time are spent on select() in a overloaded proxy server (Banga 99)

8 Interface of select() and poll()

9 Source of overhead on poll() Step by step: Read/write handler Poll() is called, array copy Linear scan in kernel Driver poll callback Poll() return, array copy Linear scan in app. Read or write handler ev_array ufds User Kernel Device driver poll()

10 Source of overhead on poll() Linear scan of ev_array and call device driver s poll callback function. Linear scan in server application to detect network events. Most work are wasted for idle connections.

11 POSIX.4 Real Time Signal Event dispatching mechanisms in Linux POSIX.4 RT extension allows signal delivery with a payload. sigwaitinfo(), sigtimedwait() Payload can carry sender PID, UID, etc. Linux extension of POSIX.4 RT signal Allow delivery of socket readiness via a particular real time signal. Pro Official support in Linux kernel since version 2.4 Con Linux specific

12 Flowchart of POSIX.4 Real Time Signal Network events Sig no FD no event type RT signal queue kernel user sigtimedwait() Read/write a FD

13 Problems in RT signals Edge-triggered readiness notification Signal queue may contain multiple events of a FD Stale event Kernel signal queue size limit entries RT signal queue overflow Some connections will fall into deadlock state.

14 Solution to RT signal queue overflow Application solution Server falls back to traditional select() or poll() Kernel solution Signal-per-fd [Chandra 01] Collapse events of the same file descriptor (signal queue size == max no. of files a process can open) => no signal queue overflow

15 Our solution to reduce poll() overhead: Scalable event dispatching library Motivation Web connections are idle most of the time. Network events in a single HTTP transaction are bursty. Our strategy Save the time wasted on idle connections for more useful works. The frequency of calling poll() on a idle connection is decreased. Average number of file descriptors at every poll() is decreased. Exploit temporal locality among events in a connection to reduce the frequency of calling poll() on a FD.

16 Live counter and polling sets: Keys to implement locality poll() A live counter is associated with a file descriptor. On every poll() to the FD: If event is detected, counter increases. If no event is detected, counter decreases. All FDs in a server are divided into three sets according to their live counter. The frequency of calling poll() on each set is different.

17 Architecture view of event dispatching library (FDM) Modular Hashing key = fd, value=nth set fdm_reg_fd(int fd) fdm_update_fd_mode(struct pollfd *fds, void *payload) fdm_wait_event() Active polling set Doze polling set Idle polling set Actual polling set sys_poll() User space Kernel space

18 Library interface design logic Reduce copy of FD array Interesting set are built gradually Separate routine to fetch event Reduce linear scan of FD array in server code Timeout timestamp and payload associated with a file descriptor is maintained in this library. Listening socket descriptor can be locked in the active polling set.

19 Proposed library interface

20 Performance evaluation Compare event dispatching mechanisms available in Linux 2.4, including our FDM library Test server dphttpd + event-threading + FDM All tests are under Multi-accept implementation Pentium III 600MHZ, 128MB RAM Test Client Httperf, HTTP workload generator three Pentium III 1GHZ, 512MB RAM

21 Dispatching overhead Goal see the overhead under Large number of idle connections Request rate is fixed at a pretty light load.

22 Server reply rate with fixed light load Reply Rate (1000 req/s, N=3, 1KB reply size) 1040 poll FDM RT signals 1020 Reply rate (reply/s) Number of idle connections

23 Server response time with fixed light load Response Time (1000 req/s, N=3, 1KB reply size) poll FDM RT signals 14 Response time (ms) FDM is more scalable than poll Number of idle connections FDM polling set maintain overhead >= polling time saved RT signal still needs to maintain timeout array

24 Dispatching Throughput Goal See the throughput under Fixed number of idle connections Overloaded request rate

25 Server reply rate with 1000 idle connections Reply rate (reply/sec) Reply Rate (1000 idle connections, N=3, 1KB reply size) poll FDM RT signals(poll) Request rate (req/sec) 100Mb Ethernet saturated, poll performs well because of multi-accept

26 Server response time with 1000 idle connections Response time (ms) Response Time (1000 idle connections, N=3, 1KB reply size) 10 poll 9 FDM RT signals(poll) Request rate (req/sec) Even at light load, we can observe overhead of 1000 idle connection. FDM suffers too since it depends on poll() Difference Between FDM And poll

27 Server reply rate with 6000 idle connections 8000 Reply Rate (6000 idle connections, N=3, 1KB reply size) Reply rate (reply/sec) RTsig+select 5000 is 4000 better than RTsig+poll, 3000 but signal queue 2000 overflow poll 1000 FDM still limit RT signals(poll) RT signals(select) scalability Request rate (req/sec) RT signal queue start overflow

28 Server response time with 6000 idle connections Response time (ms) Response Time (6000 idle connections, N=3, 1KB reply size) 200 poll 180 FDM RT signals(poll) RT signals(select) Request rate (req/sec) Signal queue overflow, server fall back to poll(), poll overhead at 6000 idle connections is a lot larger such that this behavior can t be observed at 1000 idle connections.

29 Summary of performance evaluation At light load FDM incurs low overhead At heavy load FDM improves poll() RT signal queue overflow should be protected by select(), not poll()

30 Conclusion FDM library improves poll() exploiting temporal locality property of events in a file descriptor. Better performance than RT signals if RT signal queue overflowed. RT signal queue overflow recover select() is a better choice. Availability of FDM and test web server

31 Backup slides follow

32 Keys to scalable event dispatching mechanisms Interesting set of file descriptors is built gradually inside kernel Separate building of interesting set from event retrieval Only return a file descriptor if there is a event. Collect event efficiently Cache device driver poll result Don t run driver s poll() at every poll() Or, maintain a event queue and collect events gradually at every call to TCP/IP event handler.

33 Summary of event dispatching mechanisms methods features Scalable to large set of file descriptors Event collapsing Dequeue multiple event per system call Event queue overflow When queue overflow, kernel fallback to traditional poll() Return initial state when declare interest fd Multiple interest sets maintained in kernel for per process Select() No NA Yes NA NA NA NA Poll() No NA Yes NA NA NA NA /dev/poll Yes NA Yes NA NA NA Yes RT signals Yes No No Yes No No Yes RT sig-per-fd Yes Yes No No NA No Yes Declare_interest Yes Yes Yes Yes Yes Yes No

34 Polling frequency of a set Depends on the default value of live counter Assume default value is N On every fdm_wait_event() Active polling set is polled On every N fdm_wait_event() Doze polling set is polled On every N^2 fdm_wait_event() Idle polling set is polled

35 Server performance with different default value of live counter Dispatching throughput and response time (1000 connections, 500 req/s, 1K reply size) Reply rate Response time Reply rate (replys/sec) Response time (ms) Value of live counter (N) Small N Is good Fd in Idle set Fd in doze set Fd in active set Large N useless